调研：NNAPI相关（ONNX EP）

[KakaruHayate]

尽管NNAPI已经停滞，但不考虑XNNPACK，因为目前来看支持更差。从算子支持列表来看安卓平台的EP更注重CV/NLP/LLM，比较漠视语音合成相关。不支持1D卷积算是最大障碍。

参考onnxruntime文档，以下是NNAPI的算子支持列表

Supported ops

Following ops are supported by the NNAPI Execution Provider,

Operator	Note
ai.onnx:Abs
ai.onnx:Add
ai.onnx:AveragePool	Only 2D Pool is supported.
ai.onnx:BatchNormalization
ai.onnx:Cast
ai.onnx:Clip
ai.onnx:Concat
ai.onnx:Conv	Only 2D Conv is supported. Weights and bias should be constant.
ai.onnx:DepthToSpace	Only DCR mode DepthToSpace is supported.
ai.onnx:DequantizeLinear	All quantization scales and zero points should be constant.
ai.onnx:Div
ai.onnx:Elu
ai.onnx:Exp
ai.onnx:Flatten
ai.onnx:Floor
ai.onnx:Gather	Input indices should be constant if not int32 type.
ai.onnx:Gemm	If input B is not constant, transB should be 1.
ai.onnx:GlobalAveragePool	Only 2D Pool is supported.
ai.onnx:GlobalMaxPool	Only 2D Pool is supported.
ai.onnx:Identity
ai.onnx:LeakyRelu
ai.onnx:Log
ai.onnx:LRN
ai.onnx:MatMul
ai.onnx:MaxPool	Only 2D Pool is supported.
ai.onnx:Max
ai.onnx:Min
ai.onnx:Mul
ai.onnx:Neg
ai.onnx:Pad	Only constant mode Pad is supported. Input pads and constant_value should be constant. Input pads values should be non-negative.
ai.onnx:Pow
ai.onnx:PRelu
ai.onnx:QLinearConv	Only 2D Conv is supported. Weights and bias should be constant. All quantization scales and zero points should be constant.
ai.onnx:QLinearMatMul	All quantization scales and zero points should be constant.
ai.onnx:QuantizeLinear	All quantization scales and zero points should be constant.
ai.onnx:ReduceMean
ai.onnx:Relu
ai.onnx:Reshape
ai.onnx:Resize	Only 2D Resize is supported.
ai.onnx:Sigmoid
ai.onnx:Sin
ai.onnx:Slice
ai.onnx:Softmax
ai.onnx:Split	Number of splits must evenly divide split axis size. Input split should be constant if provided.
ai.onnx:Sqrt
ai.onnx:Squeeze	Input axes should be constant.
ai.onnx:Sub
ai.onnx:Tanh
ai.onnx:Transpose
ai.onnx:Unsqueeze	Input axes should be constant.
com.microsoft:QLinearAdd	All quantization scales and zero points should be constant.
com.microsoft:QLinearAveragePool	Only 2D Pool is supported. All quantization scales and zero points should be constant.
com.microsoft:QLinearSigmoid	All quantization scales and zero points should be constant.

综上有以下结论：

1. 即时实装NNAPI，Vocoder模型（如DiffSinger的nsf-hifigan）也应该强制使用CPU推理
2. 对于DiffSinger等模型应该比较应用NNAPI前后的推理差异，频繁的fallback开销未必会小
3. 当然因为DiffSinger目前已经更倾向于使用纯MLP搭建网络，类似LLM等现代神经网络，值得一试